Introducing the Milk-V Oasis with SG2380: A Revolutionary RISC-V Desktop Experience

Oasis
142

Looks like no one is responding.

I didn’t think they would respond but tried to DM Sophgo on Twitter anyway :sweat_smile:. They haven’t responded.

Based on this post by the user with the username Sandor (who seemed to be a Sophgo employee based on his posts) the timeline they were hoping for was

Millston(we hope)：
RTL Freeze : In March 2024
Fab Tape Out: 2024-4-15 - 2024-5-30
We hope to start the super early bird pre-sale on Kickstarter and Crowdsupply platforms in March next year.

As far as I am aware, there is no early bird presale on Kickstarter/Crowdsupply. And even the latest update we got on a change to the SG2380 by @WeiMouMou was in April 2024. So they seem to be a fair bit behind their hoped schedule.

I’m not really planning to create an account on the Sophgo forum, but @mzs you might possibly get a response if you tag a (presumed) employee of theirs. Far as I can tell based on their posts the employees that I can see might be “Sandor” and “JoannaWu”.

That said, I wouldn’t want to pester them too much, so only pester them in moderation :grin:.

3 Likes
143

Been thru this before, early bird never seems to work out on schedule. But with luck, its getting closer. Really looking forward to this, finally get off the GPU treadmill.

1 Like
144

Sandor quit Sophgo months ago so no bother.
The Sophgo twitter is not updated for a while, mostly like the lay off had impact on this…

4 Likes
145

I added a single ping, I do not feel that I should ever send any more.

(Background: I once long ago, as part of a new hardware test for my job at the time, fully saturated a high capacity dedicated private intranet datalink between two physically disperse sites so that an emergency satellite backup link would kick in and handle the overflow. And command I used was the humble ping. OK it was on quite a few machines, all running multiple pings in parallel, to multiple destination machines, all sending the maximum packet size.)

2 Likes
146

That is unfortunate to hear what is happening within Sophgo and employees being laid off. Suffice to say that with the staff reduction and now with the company apparently, like everyone else in tech, also jumping onto the AI train by dedicating resources to their Dongshan super computer as well as their ARM-based AI processors to go after the high end market, does it seem more realistic SG2380 will be pushed out further till middle of next year.

If so that would be a very long delay. And since Oasis is dependent on the SoC, the board can’t start production until afterwards then either. Unless perhaps Milk-V makes adjustments to the strategy.

What about considering to modify both Oasis LPCAMM2 and LPDDR5 boards to use the compute module CM form factor with parallel Hirose connectors for the SoC interface. Since the components on the mainboard itself are common and readily available, going this route will allow production of the bare board to commence right away.

Another advantage - any other SoC could be used to test the board functionality and ports as well. With that said, a good candidate for a compute module would be the 8-core Eswin EIC7702x that very similarly, is SiFive-based:

https://www.eswincomputing.com/en/products/index/36/10.html

Add a dual-channel capable memory controller along-side the EIC7702x onto a compute module, that would complete the pieces needed for a bootable linux system to validate Oasis hardware.

In fact, Eswin supposedly reported their manufacturing lines are already up and running for mass production. If true and Eswin can actually commit to large quantities, that even opens up the feasibility to go forward with the Oasis launch and pivot to bundle the EIC7702 CM with the board. Then at least people can get their hands on actual hardware to start some bit of development or initial testing and setup in the interim while waiting until SG2380 finally comes out, which should be a simple drop-in replacement.

Or who knows, Eswin could come through with their own 16-core EIC variant in the near future as well. (How about 16x P870-D cores for the next iteration, and maybe finally come to their senses to dump Imagination chips to go instead with VeriSilicon Vivante GC7000L iGPU and make it even more compelling…)

Though most everyone that were enticed by news of Oasis are probably looking forward to the 16-core and holding out for that to be released. However if it is seen possible to replace the SoC and upgrade in the future, likely the interest would still remain from everyone here to buy the new bundle with 8-core instead? It would still be useful since the other important board features such as PCIe slot and 4x SATA that also got attention would be immediately available.

Also Oasis might be a good base to build a long term product line around. Having a CM interface allows such a case. If offers flexibility to respond to the rapidly changing RISC-V landscape and enables the ability for end users to try out the latest and greatest hardware more easily. Instead of wasting time having to start from scratch each time fighting with things like nvme or sata for each new board in order to use it, more effort could be put into focusing on the actual RISC-V part of the hardware like getting the new vector extensions optimized into gcc and move the arch further along more quickly.

The mainboard already has the primary interfaces that people are interested in for a variety of different uses. All Milk-V has to do is to produce additional future compute modules with whatever new SoC comes out, which should be simpler to set up manufacturing with it having a smaller footprint. On the consumer side, seeing stable hardware with signs of longevity from the manufacturer instills trust and brand loyalty. Many people do not understand that is one of the reasons for the popularity of Raspberry Pi. The Pi ver1 is still getting updates with the Pi team continuing to release a usable image of the newest OS that works on all their old hardware. That shows commitment from the vendor and that people’s investment were worth it and to continue using their hardware.

2 Likes
147

“The AI Train”. Well that IS the market, its not some fad like many clueless people think, like it or not. So of course that is where things are designed for. But along side the data center level products, ‘edge and consumer’ AI is also a market, however less profitable so it will get pushed down a bit on importance i do agree.

But if Oasis gets pushed back too far, someone else will fill that need, and may not even be worth continuing.

And as a side note, RPI. sucks, is a joke, and is for children. But i wont get into that here. Different market. Different needs. Comparing apples to rocks. ( not even oranges )

1 Like
148

Probably work out a lot cheaper to modify the Megrez board design to support an EIC7702X SoC and call it the Megrez 2.

1 Like
149

Is the Oasis actually officially delayed? There is not much Q3 left and I see no announcements.

1 Like
150

I would say that Milk-V really do want to release the Oasis which based around a SG2380 System on a Chip, that will eventually be produce by Sophgo. The problem is that Sophgo have not yet officially finalized the tape out for the SG2380 SoC and send it off to TMSC to manufacture a physical microchip that can be held in your hand (and a few posts back mentioned that there were some involuntary redundancies at Sophgo, which is never good for product deadlines - let alone the lives of the people affected). So I would say that Milk-V Oasis is flying in a holding pattern waiting on the Sophgo SG2380 to land before their motherboards can hit the runway and takeoff. I do not think that it has been officially stated, but it is impossible to deliver a ITX motherboard based around a SoC which does not exist as a physical product that can be bought (yet).

2 Likes
151

So it does exist as spotted by sharp eyes and was not a mirage after all! But to get Megrez up to feature parity to Oasis with the hardware that got people’s attention? It is unlikely that modifying the Megrez will be cheaper. Having to make minimal and least costly changes was part of the consideration with the suggestion for going to the CM form factor. To enable the compute module for Oasis, only 2 inexpensive connectors are needed and to redo the traces around the current SoC area.

As mentioned the appeal for Oasis was not only with the 16-core processor, but also the board features like 4x SATA, 3x M.2 (which includes NVMe PCIe), eDP, MIPI CSI, LPCAMM2 that are currently not there on Megrez. Those are a lot more components to add, and rearrange existing electronics around for the necessary controllers and physical interfaces. On top of that it requires retracing throughout the entire board.

Also by adding these features to the Megrez, wouldn’t that make it - the Oasis board…

1 Like
152

Looks like you got a reply.

Though the account that replied… maybe I’m overthinking but the profile picture it’s using seems very AI generated for some reason. Even the content of it’s reply is something I’d expect ChatGPT to spit out if I asked it “Explain reasons for why a chip (SG2380) might have a delayed tapeout”.

The SG2380’s tape out status is mostly determined by a variety of factors; including design difficulties and testing stages. It is critical to understand that delays are common in chip creation due to the intensive testing required to ensure operation.

1 Like
153

Just sounds like a standard " we don’t know " reply without saying they don’t know.

Consolation is sounds like its not cancelled or anything.

154

That’s just a random user that’s trolling.

155

I’m not sure if it’s just some user trolling. At least this isn’t the first time I’ve come across accounts on a forum where the profile picture seems subtly AI generated and the comment itself seems to be generated by an LLM.

It might be a user trolling. Or it might be some company/research team mass-generating AI accounts for various forums and getting the AI to interact with people and using the interactions as more training data to create an LLM that can more accurately mimic human conversation.

Maybe too conspiratorial. It might just be Sophgo themselves setting up an AI account to reply to folks on their forums so they can reduce their staff.

1 Like
156

My initial thought was the next generation of large language model enhanced spambots. create an account, select recent random thread. use the last post make a human like reply. Archive spambot account for several weeks, post spam.

Problem is forums with human moderators can spot spam and remove spam in seconds.

Checkout the Orange Pi forums (there appears to be no mods there) for an example of spambots running wild there are ten to twenty spambot messages to every one human post. Spambots responding to spambots spam can get funny fast.

3 Likes
157

Checking around the Sophgo forum, the response does not appear to be from an employee as indicated in the other post listed at the top

https://forum.sophgo.com/t/i-need-some-advice-on-optimizing-ai-model-performance-with-sophgo-hardware/709

So still no official company response regarding SG2380. Though the comment is true that technical issues may crop up to delay a product, there are background info in this particular case that was probably not aware of. Namely what is concerning here is it appears that even Milk-V, a hardware partner that already successfully sold a lot of boards with their other processors, has not been kept up to date on the status either. Along with the news Sophgo laid off people but still released other AI server products this year, so may be distracted pursuing the high end market.

The same thing happened to SiFive. Intel was supposed to mass produce the latest and greatest SoC and HiFive Pro P550 dev board meant to be sold last year:

https://www.hackster.io/news/sifive-s-hifive-pro-p550-developed-in-partnership-with-intel-aims-to-be-the-fastest-risc-v-sbc-yet-5789a4734bcc

But besides a few Alpha boards, Intel did not manufacture any finished product in volume nor release to the public for purchase. Now, look at Intel and all their troubles, coming clean about huge losses and cut backs while HiFive Pro is no longer listed on the websites any where. Instead out of the blue, SiFive recently announced the new HiFive Premier P550, a year later than planned. Notice the new board went with a modular CPU this time, without Intel. But still leaves room for an Intel branded SoC in the future just in case.

Having to revamp a product is a key take away from SiFive’s experience. Just goes to show that well established companies are not immune to set backs, and even a big name like Intel can falter on their promises. Simply the reality of the manufacturing business that many do not understand.

Through no fault of their own because third party internal affairs and events are beyond their control, Milk-V may be looking to be in the same situation. However contrary to what some may believe, Milk-V can take matters into their owns hands and explore alternative solutions to keep moving forward just like SiFive.

As suggested a possible option to change tactics is positioning Oasis as a hardware platform, and not be dependent on a specific SoC. The result is Oasis boards without the SoC could start to be produced right away and yield some engineering samples to start the testing sooner. SiFive finally got wise in the end but if they had designed modularity from the start would not have been set back a whole year.

But the general specs for the Oasis board at the affordable price as originally planned is still compelling. Unfortunately initial SBC hardware did not fulfill the basis for a good RISC-V platform to its full potential, primarily due to Imagination whom everyone had given the benefit of the doubt in good faith but another vendor only delivering lackluster promises. Case in point there is a reason even 2+ years after launch there is still no official Ubuntu Desktop for JH7110-based boards. Only Ubuntu Server can be used, still requiring a serial console for setup at that. However people have shown RISC-V boards being able to run x86 Steam games given the right hardware. What could have been if SoC manufactures had gone with the Vivante GC7000L in retrospect - probably a lot more stable full featured images already from many others like Arch, Alpine, Gentoo, Manjaro, etc, instead of partial workinig ports everywhere while trying to reverse engineer hardware not related to RISC-V architecture. Another set back.

So it would be greatly beneficial for the RISC-V ecosystem to have access to Oasis hardware which more people can afford and drive development further. The growth and wide spread usage of an open platform is the end goal for everyone here interested in the product. Not sure when the last direct communication Milk-V received from Sophgo, but if it was recent, then it may make sense to give some additional period for an official respond. On the other hand if it has been a long while already with no answers, it may be for their own best interest to initiate other solutions soon if Milk-V still have faith in Oasis as a viable product.

4 Likes
158

You’re right. Instead of randos like us tweeting at Sophgo or posting on their forums, I would have thought it’s much more likely for MilkV to get a response from them directly; they are a customer of theirs after all. Milk-V probably has some internal contact at Sophgo/can call them and ask about it.

On the topic of GPUs. Yes, Imagination is extremely disappointing. But Vivante doesn’t seem amazing either? Checking the specs for the GC7000L, it might be able to match the BXE-4-32, but is by far outclassed by the AXT-16-512 that we are currently expecting on the Oasis. Secondly, I might be mistaken, only the GC8000 series cards have some support for Vulkan (1.2 max I think so no Vulkan 1.3) and no open drivers for Vulkan. The GC7000L doesn’t support Vulkan at all it seems? Lastly, looking at mesamatrix for a rough overview of the Vivante driver state, it seems to be in the worst state among all the OpenGL drivers?

2 Likes
159

Milk-V is not getting answers from their contacts since they already mentioned they do not know the current status either, which is the issue.

The reason for bringing up GC7000L is because the linux driver is mature and fully open source

https://www.phoronix.com/news/Etnaviv-2020-Driver-Progress

https://www.phoronix.com/news/Etnaviv-Driver-Linux-6.11

Being unencumbered by any proprietary blob means an easier time to bring up a full distro quickly and would have been useful back then when RISC-V silicon that were available to end users first appeared.

It is the same deal when comparing between AMD and Nvidia among graphics cards. Radeon RX are working under RISC-V. Why? Because drivers are 100% open source. No mention of Nvidia on RISC-V anywhere to be found. Nvidia still wants total control over their drivers, even though they started contributing to mainline linux. But RISC-V won’t happen until they release the binary bits themselves for the arch (or if people reverse engineer each GPU firmware and good luck not getting sued by them as they aggressively guard their CUDA secrets). No reason then to bother with the headache of Nvidia right now. AMD works on RISC-V, x86, ARM - any hardware people decides to use as long as linux runs and PCI implementation is up to spec. That is the power of open source.

So all the great GPU specs on paper does not help if Imagination locks those features behind driver blobs that are tied to specific, older kernels, ON PURPOSE. They are acting like Nvidia.

Look at the situation with Ubuntu as mentioned. How good are those specs and features of Imagination’s GPU when a monitor can not be even used with a board. The latest Ubuntu Desktop should be working, but no. They are unwilling to provide updated drivers nor submit the needed code themselves into mainline nor release any tech docs without an NDA so the community has the option to do the work for them. It seems fair to say Imagination has had ample opportunity to provide any solution and should have happened yesterday already.

It is obvious they are binding their time in hopes that Google can untangle the convoluted hole they dug themselves into to finally get Android RISC-V out the door soon. Then revert back to their preferred operation as before where they can just drop off a proprietary driver for a single version of a phone and then wash their hands of any further support. What is the point to keep using their product then if the situation will be the same for any future hardware. Isn’t it best to drop unusable baggage and switch gears to bring other competition into the space.

Obviously the ideal RISC-V SoC would be to have AMD on the die. Just imagine, a nice little RISC-V SBC with RDNA clusters inside pushing pixels. But that is fantasy. No way AMD would license their iGPU for use in an ISA that is a competitor to them. All other existing open source candidates on the list, like AMD, are not licensible IP either with the remaining choice being GC7000L. It may not be ideal but having some graphics is better than paying for hardware that does not work while staring at a blank screen, especially dealing with a small SBC. At least for larger boards there is the ability to include PCIe slots to upgrade to better hardware, or even possibly to embed a discrete GPU like the AMD RX 550 on the mainboard.

But options being what they are at the moment beggars can’t be choosers at this early stage. Getting into a usable state as quickly as possible to dive into optimizing all the software for the RISC-V side of hardware will go much faster without first having to fudge around with components like the GPU and other undocumented interfaces. How many users abandoned their RISC-V SBC because something simple like video output does not function on the hdmi port that came with the board. When development for a new arch is moving at such rapid pace what is important are known components working with expected functionality, and that means fully open source software under linux to allow developers the flexibility and freedom to forge ahead without artificial restrictions.

2 Likes
160

Well, I guess I would agree that it would have been useful for the earlier SoCs and SBCs to use Vivante’s GPUs like the JH7110 and the SpacemiT K1. You’d probably be able to easily get GPU support in mainline on them.

But again, Vivante isn’t amazing. AFAICT it’s not even GLES3 so far?

I have mixed feelings about the IMG AXT-16-512 on the Oasis. Yes, it is a lot more powerful than Vivante’s stuff. But it has a different architecture from it’s other GPUs released so far. So they will need to do a lot of work to bring driver support to it. I got the feeling we might not get driver support until 2025H2 or even 2026.

Well, I’m guessing/hoping by 2025 H1 the IMG GPUs released so far will have much better driver support because IMG folks seem to say they’ll be done with driver work for the AXE-1-16M and BXS-4-64 by end of year 2024.

2 Likes
161

Yes, a 2 year set back, and what could have been. Yet even presently, Imagination could have this fixed right now with Ubuntu and other distros if they wanted to. So still want to trust them? Even while their linux driver they are working on continue to require binary blobs, so just the bare minimum to claim “open source” like Nvidia. How much further set back should this go on…

The suggestion for GC7000L was not specifically for Oasis. It may be too late to uncommit to Imagination if there is the possibility to use an another existing SoC and start manufacturing soon. In any case for Oasis, the iGPU would not matter because of the PCIe slot. Likely for most people it will be populated with an AMD card. In fact for whatever SoC to be used in Oasis, it even could go without an iGPU. However that does not help an SBC without some type of expansion.

But going forward best to switch to GC7000L. If the SoC is used in a credit-card sized SBC, at least there is some graphics. On larger boards like Oasis the same SoC can be used but the PCIe slot is available if needing better graphics for the use case. Or go with an embedded discrete GPU like the Radeon RX 550 on the mainboard but that will increase the price of the board at bit which could be acceptable for some.

Is that not better than a useless hdmi port and being abandoned hardware? Many people will avoid and not buy such a RISC-V board otherwise when finding out it includes Imagination, no matter what the GPU specs are. Which seems like that may be starting to happen based on comments around other forums now their reputation is becoming wider known. There is even a sense the hardware vendors are becoming aware of this too. Notice how with many new RISC-V announcements, the identify of the GPU is not given and only described in general terms like “super duper 3D graphics”. That is going to hurt RISC-V adoption and delay further.

The primary appeal and promise of RISC-V is an open hardware platform, but the example for the current situation with Ubuntu indicates it is not. What good is having Vulkan in that case. How is that usable. Why keep hoping and stay the course when something can be done and there exists solutions for a true FOSS RISC-V stack. Build up and improve from there. Especially considering the current vendor could have this resolved already with recompiled drivers or needed information but is choosing not to. So how long to wait to amputate the leg before realizing it’s too far gone and the danger of tumor spreading to the remaining good parts.

3 Likes